Is it possible to get the observed information matrix (i.e. the Fisher matrix, but just for every sample in a minibatch separately, not the expectation over the data distribution) by using your code? If yes, could you perhaps give a rough outline on how to accomplish this? Thanks!